Research map / Architectures and efficiency / CNN design
Clustering Convolutional Kernels to Compress Deep Neural Networks
Sanghyun Son, Seungjun Nah, Kyoung Mu Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase number of centroids k from 128 to 1024 (ablation) | VGG-16 kernel clustering · CIFAR-10 test | error rate (%): 6.24 -> 6.01 | Improved |
| increase centroids from 128 to 1024 (ablation) | ResNet-56 kernel clustering · CIFAR-10 test | error rate (%): 6.76 -> 6.30 | Improved |
| add multi clustering with 2 groups (ablation) | VGG-16-C64 · CIFAR-10 test | error rate (%): 6.44 -> 6.27 | Improved |
| add per-layer multi clustering (M13) (ablation) | VGG-16-C64 · CIFAR-10 test | error rate (%): 6.44 -> 6.21 | Improved |
| add per-layer multi clustering instead of 2 groups at k=256 (ablation) | VGG-16-C256 · CIFAR-10 test | error rate (%): 6.16 -> 6.10 | Improved |
| add transform invariant clustering with vertical flips (TIC2) (ablation) | VGG-16-C128 · CIFAR-10 test | error rate (%): 6.23 -> 5.92 | Improved |
| add horizontal flips too (TIC4, 64 centroids) (ablation) | VGG-16-TIC2 · CIFAR-10 test | error rate (%): 5.92 -> 6.25 | Hurt |
| add rotations too (TIC8, 32 centroids) (ablation) | VGG-16-TIC · CIFAR-10 test | error rate (%): 6.25 -> 6.51 | Hurt |
| add filter pruning before clustering (ablation) | VGG-16-C512 · CIFAR-10 test (with pruning) | error rate (%): 6.16 -> 5.99 | Improved |
| add simple pruning before clustering (ablation) | VGG-16-C512 · CIFAR-10 test (with pruning) | error rate (%): 6.16 -> 6.13 | Improved |
| remove scale parameters (N) at 128 centroids (ablation) | DenseNet-BC-C128 · CIFAR-10 test (DenseNet-BC) | error rate (%): 4.51 -> 4.44 | Improved |
| remove scale parameters (N) at 64 centroids (ablation) | DenseNet-BC-C64 · CIFAR-10 test (DenseNet-BC) | error rate (%): 4.50 -> 4.60 | Hurt |
Closest papers by meaning
- Exploiting Kernel Sparsity and Entropy for Interpretable CNN Compression (CVPR 2019 · similarity 0.91)
- Deep k-Means: Re-Training and Parameter Sharing with Harder Cluster Assignments for Compressing Deep Convolutions (ICML 2018 · similarity 0.90)
- DKM: Differentiable k-Means Clustering Layer for Neural Network Compression (ICLR 2022 · similarity 0.88)
- Holistic CNN Compression via Low-Rank Decomposition with Knowledge Transfer (TPAMI 2018 · similarity 0.87)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.85)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
